AI UX glossary
Plain-English definitions of AI product terms for UX designers, PMs, and product teams: how models behave, where they fail, and what to design around them.
Foundations
- Foundations
AGI (Artificial General Intelligence)
AGI (artificial general intelligence) is the idea of AI that matches or exceeds human ability across most cognitive tasks, not just one narrow skill like translation or image generation.
- Foundations
AI Product Design
AI product design is the practice of shaping how people understand, trust, and control AI-powered features—not only how they look.
- Foundations
AI UX
AI UX (AI user experience) is how people perceive and interact with AI features: streaming replies, tool pickers, citations, autonomy, memory, and recovery when something goes wrong.
- Foundations
Artificial Intelligence (AI)
Artificial intelligence (AI) is software that performs tasks that normally require human judgment, such as understanding language, recognizing patterns, making recommendations, or generating content.
- Foundations
ASI (Artificial Superintelligence)
ASI (artificial superintelligence) is hypothetical AI that surpasses the best human minds across virtually all domains, including scientific creativity and strategic planning.
- Foundations
Design Engineer
In AI product teams, a design engineer sits between design and code—prototyping AI flows in real stacks, defining motion and states agents must preserve, and reviewing generated UI for trust and craft.
- Foundations
Deterministic vs. Stochastic
Deterministic systems return the same output for the same input every time. Stochastic (probabilistic) systems, including most LLMs, sample from possible answers so results can vary run to run.
- Foundations
Diffusion
Diffusion models generate images (and sometimes video) by iteratively refining random noise into coherent visuals guided by a text prompt.
- Foundations
Large Language Model (LLM)
A large language model (LLM) is an AI model trained on vast text to predict and generate language, used in chat assistants, copilots, search answers, and agents.
- Foundations
Machine Experience (MX) Design
Machine experience (MX) design is designing products so both humans and AI systems can read structure, meaning, and relationships—not only pixels.
- Foundations
Machine Learning
Machine learning (ML) is a branch of AI where systems improve at a task by finding patterns in data instead of being explicitly programmed for every scenario.
- Foundations
Model
A model is the trained engine that turns inputs (text, images, context) into outputs (answers, labels, actions). ChatGPT, Claude, and Gemini are products built on large language models.
- Foundations
Multimodal
Multimodal AI can process and generate more than text: images, audio, video, and structured files in the same workflow.
- Foundations
Probabilistic UI
Probabilistic UI is interface design for software that can produce different outputs for the same input—LLMs, recommenders, generative media—not deterministic CRUD.
- Foundations
RLHF (Reinforcement Learning from Human Feedback)
RLHF aligns a model’s behavior with human preferences by training on rankings, ratings, or corrections from people, not just raw internet text.
- Foundations
Token
A token is a small chunk of text (often a word fragment) that models use to read and write language. Input and output size, cost, and speed are measured in tokens.
Prompting and interaction
- Prompting and interaction
Chain of Thought
Chain of thought (CoT) is a prompting approach where the model shows intermediate reasoning steps before the final answer.
- Prompting and interaction
Few-Shot Prompting
Few-shot prompting includes a small number of example input/output pairs in the prompt so the model mimics your format, tone, or decision style.
- Prompting and interaction
Multimodal Input
Multimodal input lets users combine text with images, audio, video, or files in one request so the model can reason across media types.
- Prompting and interaction
Progressive Disclosure
Progressive disclosure reveals AI output in layers: summary first, details on expand, advanced controls only when needed.
- Prompting and interaction
Prompt
A prompt is the instruction or question you give an AI model: the user message, template, or form that tells it what to do.
- Prompting and interaction
System Prompt
A system prompt is the hidden instruction layer that defines the AI’s role, rules, tone, and boundaries before any user message appears.
- Prompting and interaction
Zero-Shot Prompting
Zero-shot prompting means asking the model to perform a task with instructions only, with no examples of desired input/output pairs.
Retrieval and model behavior
- Retrieval and model behavior
Context Window
The context window is the maximum amount of text (in tokens) a model can consider in one request: your prompt, system instructions, retrieved docs, and chat history combined.
- Retrieval and model behavior
Embeddings
Embeddings are numerical representations of meaning that let systems compare how similar two pieces of text (or images) are, even when wording differs.
- Retrieval and model behavior
Grounding (UX)
Grounding in UX is how an interface ties AI answers to verifiable sources—documents, URLs, files, or tool results—so users can check claims.
- Retrieval and model behavior
Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation (RAG) retrieves relevant documents or records first, then asks the model to answer using that material.
- Retrieval and model behavior
Semantic Search
Semantic search finds content by meaning and intent, not just exact keyword matches, powered by embeddings and vector comparison.
- Retrieval and model behavior
Vector Database
A vector database stores embeddings and retrieves the nearest matches quickly: the infrastructure behind semantic search and RAG at scale.
Safety and trust
- Safety and trust
Capability Disclosure
Capability disclosure is the practice of telling users—in plain language—what an AI feature can and cannot do before and during use.
- Safety and trust
Explainability
Explainability is how clearly an AI product shows why it produced an answer: sources, reasoning steps, confidence, or feature influence.
- Safety and trust
Guardrails
Guardrails are rules, filters, and policies that block unsafe inputs, limit risky outputs, and keep AI behavior aligned with product and brand standards.
- Safety and trust
Hallucination
A hallucination is when an AI states something confidently that is false, outdated, or unsupported by its inputs.
- Safety and trust
Human-in-the-Loop
Human-in-the-loop (HITL) means a person reviews, approves, or corrects AI output before it affects users, records, or systems.
- Safety and trust
Moderation
Moderation is the process of detecting and handling harmful, abusive, off-brand, or policy-violating content in AI inputs and outputs.
- Safety and trust
Prompt Injection
Prompt injection is when untrusted text tricks the model into ignoring its instructions or exposing hidden system behavior.
- Safety and trust
Slop (AI Slop)
AI slop is low-quality, generic, or misleading AI-generated content flooding feeds, search results, and products: templated articles, fake thumbnails, and plausible but empty copy.
Agents and workflows
- Agents and workflows
A2A (Agent-to-Agent)
A2A (agent-to-agent) communication is when one AI agent delegates tasks, shares context, or negotiates outcomes with another agent instead of only talking to the user.
- Agents and workflows
Action Receipt
An action receipt is a plain-language record of what an agent did, when, and with what permissions—like a bank notification, not a server log.
- Agents and workflows
Agent
An agent is an AI system that plans steps, uses tools, and acts across multiple turns to complete a goal, not just answer a question.
- Agents and workflows
Agentic UX
Agentic UX is the design of interfaces for software that plans and acts on a user's behalf—tools, files, APIs, and multi-step workflows—not only text replies.
- Agents and workflows
Autonomy Slider
An autonomy slider lets users set how independently an agent may act—from suggest-only to draft to execute—often per task or surface.
- Agents and workflows
Claw
In agent tooling, a claw is the reach of an AI orchestrator into your environment: spawning sessions, running commands, editing files, or driving browsers on your behalf.
- Agents and workflows
Harness
A harness is the runtime environment that wraps a model with tools, policies, memory, and orchestration so agents can act inside your product or repo.
- Agents and workflows
Intent Preview
Intent preview shows what an agent plans to do before it executes—especially for multi-step or irreversible actions.
- Agents and workflows
Loop Engineering
Loop engineering is designing iterative cycles where an AI observes context, acts, checks results, and repeats until a task completes or a human stops it.
- Agents and workflows
MCP (Model Context Protocol)
MCP (Model Context Protocol) is an open standard for connecting AI assistants to external tools, data sources, and services through a shared protocol.
- Agents and workflows
Orchestration
Orchestration is how an AI product coordinates multiple models, tools, retrieval steps, and human checkpoints into one coherent run.
- Agents and workflows
Responsive Salience
Responsive salience is when an AI interface automatically increases or decreases oversight UI—explanations, approvals, transparency—based on task risk and user context.
- Agents and workflows
Tool Use
Tool use is when a model invokes external functions (search, calculators, APIs, code runners) to go beyond plain text generation.
- Agents and workflows
Workflow
An AI workflow is a multi-step process that combines prompts, tools, human review, and structured handoffs to reach an outcome, not a single chat reply.
Output and formats
- Output and formats
AI-Native Design System
An AI-native design system encodes tokens, components, and rules that both designers and coding agents can consume—often via DESIGN.md, SKILL.md, or structured exports.
- Output and formats
Design.md
Design.md is a markdown file that captures design system rules, component usage, tokens, and UX standards for humans and AI agents building your product.
- Output and formats
Generative UI (GenUI)
Generative UI (GenUI) is when AI produces live interface elements (forms, dashboards, cards, charts) from prompts or data, not just static text.
- Output and formats
JSON
JSON (JavaScript Object Notation) is a lightweight text format for structured data: keys, values, arrays, and nested objects that machines parse reliably.
- Output and formats
Markdown (MD)
Markdown is a plain-text format for headings, lists, links, and code that stays readable to humans and easy for models to generate.
- Output and formats
Schema
A schema defines the shape of data: required fields, types, allowed values, and relationships between properties.
- Output and formats
SKILL.md
SKILL.md is a markdown file that teaches coding agents (Cursor, Claude Code, Codex) how to behave on repeatable tasks: rules, steps, and quality bars teams install once.
- Output and formats
Structured Output
Structured output means the model returns data in a predictable schema (JSON, tables, form fields) instead of free-form paragraphs.
Product and performance
- Product and performance
AI Evals (Evaluations)
AI evals are automated or human frameworks that measure model accuracy, bias, safety, and task performance before and after you ship.
- Product and performance
Compute
Compute is the processing power (GPUs, TPUs, cloud instances) used to train models and run inference when users generate, classify, or embed content.
- Product and performance
Fine-Tuning
Fine-tuning adapts a base model to your domain, tone, or task by training on curated examples, beyond what a system prompt alone can reliably enforce.
- Product and performance
GEO (Generative Engine Optimization)
GEO (generative engine optimization) is the practice of shaping content and structure so AI answer engines (ChatGPT, Perplexity, Gemini, Claude) cite and summarize your product accurately.
- Product and performance
Inference
Inference is running a trained model on new inputs to produce outputs: the live “prediction” step users experience as chat, classify, or generate.
- Product and performance
Latency
Latency is the delay between a user action and a usable AI response: time to first token, time to complete answer, or time to finish an agent run.
- Product and performance
Memory
Memory is how an AI product retains user preferences, facts, or past context across sessions, beyond the single context window.
- Product and performance
Personalization
Personalization tailors AI behavior or content to a user or segment, using memory, history, embeddings, or fine-tuned priors.
- Product and performance
Streaming Response
Streaming is when the AI sends its answer incrementally as tokens generate, instead of waiting for the full reply.
- Product and performance
Token Burn Rate
Token burn rate is how fast a product consumes tokens over time: per request, per user session, or per agent run.
- Product and performance
Vibe Coding
Vibe coding is iterative building with AI coding tools (Cursor, Claude Code, Lovable, etc.) where natural language steers rapid prototypes—“make it feel calmer,” “add citation chips.”
Frequently asked questions
What is the AI UX glossary?
The AI UX glossary defines 70 practical terms for designers, PMs, and marketers building AI products (LLMs, RAG, agents, guardrails, latency, and more) in plain English with UX-focused examples.
Who is the glossary for?
Product designers, UX researchers, PMs, and marketers who need shared vocabulary for specs, critiques, and AI feature reviews without reading ML research papers.
How does the glossary relate to patterns and frameworks?
Glossary terms explain concepts; patterns show interface conventions with demos; frameworks organize territories like agentic UX or chat UX. Term pages link to related patterns, prompts, and frameworks when a concept maps to shipped UI.
How should I use glossary terms in product work?
Use definitions in PRDs and design reviews, link term pages in specs for alignment, and follow related patterns when a term implies interface requirements: citations for RAG, approval steps for agents, streaming affordances for latency.